Krea 3 Explained: Multi-Image Editing, Pricing, Benchmarks and the ComfyUI Node

Krea 3 is Krea’s new image editing model. You give it up to eight reference images and an instruction, and it returns a native 2K image, usually in under 8 seconds.

Krea published a full technical report on 3 November 2026. Its Ultra version sits at #4 on three public image-editing leaderboards, and it costs a fraction of Nano Banana Pro or GPT Image for the same work.

One thing to know up front: Krea 3 is a hosted model. There are no open weights, so it does not run on your own GPU. You use it in the Krea app, through the API, or inside ComfyUI through the official Krea 3 Image partner node.

This guide covers what Krea 3 does, every node setting, real prices per image, the benchmark ranks, how it was built, its weak spots, and fixes for common ComfyUI problems.

What Krea 3 Is

Krea 3 is an image-to-image model built on the Krea 2 pretrain. Its job is editing: keeping a character or product consistent, making a targeted change, or combining several reference images into one scene.

If you have used Krea 2 locally, it is a different product. Krea 2 has open weights you can run in ComfyUI. I covered it in my Krea 2 BBOX guide and the Krea 2 Turbo 3-step LoRA guide. Krea 3 is cloud-only, and the report says nothing about releasing weights.

What it can do:

• Multi-image editing: up to 8 references in one generation, such as a person, an outfit, a product and a location.
• Precise edits: change one area while keeping the rest of the image pixel-identical.
• Text-to-image: leave the image inputs empty and it generates from the prompt.
• Native 2K output, or 1K when you want it cheaper.

Krea 3 Fast vs Ultra vs Ultra Thinking

In ComfyUI these map to the node’s thinking_mode setting: fast, high and ultra. More on that below.

Use Krea 3 in ComfyUI

The Krea 3 Image node is an official partner (API) node. It was added in ComfyUI PR #17067 on 15 October 2026.

Setup

1. Update ComfyUI to a build from after mid-October 2026. On the portable build, run update\update_comfyui.bat.
2. Sign in to your Comfy account inside ComfyUI and add credits. Partner nodes bill per image from those credits.
3. Add the node: double-click the canvas and search “Krea 3 Image”, or browse to partner/image/Krea.
4. Connect images in order to the image inputs, write your instruction, and run.

There are no model files to download and no VRAM needed, because the work happens on Krea’s servers. Even a laptop without a GPU can run it.

Referencing images in the prompt

Images are numbered by the input they are plugged into. Refer to them as @Image1, @Image2 and so on:

Put the woman from @Image1 in the jacket from @Image2, standing in the cafe from @Image3, soft morning light, keep her face and hair the same

Leave every image input empty and the node does plain text-to-image.

Every Node Setting Explained

Two rules about ultra catch people out. It needs a reference image, and it always outputs 2K. It also does not do precise edits. Use high or fast with precise_edit for small fixes.

Pricing per Image

In ComfyUI, billing is per image, plus a surcharge for each connected reference. Approximate rates:

For example, a 2K fast edit with three references comes to about $0.086 + 3 x $0.03 = about $0.18. Krea’s own pricing page lists $0.06 per image in the Krea product.

At volume, Krea’s report compares monthly cost for 100,000 edits at API prices:

These are Krea’s figures, so take the comparison as their claim. Prices change often, so check the Krea 3 page before planning a budget.

Speed vs Nano Banana Pro and GPT Image

Krea ran single-reference 2K edits through the fal API:

The report says its fastest setup takes about 4 to 5 seconds, and Krea is honest that it is “not yet real time”. Most of the delay comes from two places: rewriting your prompt with an in-house vision-language model, and the diffusion sampling itself.

Benchmark Ranks

Krea calls itself the “#2 image generation lab in the world” on the strength of this. That is their own reading of the boards. What is clear is that it competes with the big labs while being much faster and cheaper.

How Krea 3 Was Built

Krea calls the architecture “deliberately standard”. The interesting work is in training and data.

• Base: the Krea 2 pretrain, continued on a larger text-to-image set that includes native 2K images.
• References as clean tokens: each reference image is VAE-encoded and added to the DiT’s sequence as clean tokens at timestep 0. Because they never change, their key-value pairs can be cached, which speeds things up.
• Text encoder: a vision-language model looks at the references at 384×384 alongside your instruction. Only the instruction tokens are passed on to the DiT.
• Training mix: 30% text-to-image and 70% image-to-image. Reference and target resolutions are sampled separately, so a 2K output can come from small references.
• Data: edit pairs mined from billions of images. Only 0.17% of candidate pairs survived all the filters. There are also synthetic renders (such as edited website HTML and 3D scenes) and computer-vision transforms like depth and Canny edges.
• Reinforcement learning: 16 rollouts per input, scored by preference judges, a defect judge, a text reader, a typography judge and a pixel-level region reward. Krea reports about 25% fewer text errors on held-out benchmarks.
• DPO: a final human-preference stage that improved texture, colour, brightness and facial identity.

Limitations

• Dense, tiny text is still imperfect. Large headlines are fine, but fine print on labels can break.
• Identity drifts over chained edits. Edit an edit of an edit and faces slowly change. Krea plans to lean on LoRAs for specific characters and products.
• Some outputs copy a reference too closely instead of changing it.
• No local option. There are no weights, and every image costs credits.
• Seeds do not reproduce, because of the prompt rewrite.

Prompt Tips for Multi-Image Edits

• Name each image’s role: “face from @Image1, outfit from @Image2, background from @Image3”.
• Say what must not change: “keep the logo and label text exactly the same”.
• Edit from the original each time instead of chaining, to avoid identity drift.
• Use precise_edit for small fixes such as changing a shirt colour.
• Draft at 1K fast, then rerun your favourite at 2K or ultra.

If you want a fully local alternative for consistent characters, my Anima Reference Adapter guide covers an open option for anime styles.

Troubleshooting

I cannot find the Krea 3 Image node

Your ComfyUI is older than the mid-October 2026 builds. Update it and restart. Also check that API nodes are not disabled in your settings.

Unauthorised or “insufficient credits” error

Sign in to your Comfy account inside ComfyUI and add credits. Partner nodes do not use your own Krea subscription.

Ultra mode fails with no image

Ultra needs at least one reference image. Connect one, or switch to fast or high for text-to-image.

precise_edit errors out

Precise edit needs at least one connected image, and it does not work in ultra mode.

The same seed gives a different image

That is expected. Krea rewrites the prompt before generating, so seeds do not reproduce results exactly.

It used the wrong reference

Check the order of the image inputs. @Image1 is the first connected input. Name each image’s role clearly in the prompt.

My Take

Krea 3 is a strong pick if you edit a lot of product or character images and care about speed and cost. Being about three times faster than Nano Banana Pro for top-five quality is a real selling point.

For me, the ComfyUI node is the most useful part. You can drop it into a local workflow, use it only for the hard multi-image step, and keep everything else on your own GPU. Just remember that each run costs credits and nothing runs offline.

FAQ

Is Krea 3 open source?

No. Krea has not released weights for Krea 3. It is available through the Krea app, the API and the ComfyUI partner node.

Can I run Krea 3 locally on my GPU?

No. The ComfyUI node sends the job to Krea’s servers. For local editing, look at open models like Krea 2 or Anima.

How much does Krea 3 cost?

In ComfyUI, about $0.043 to $0.372 per image depending on mode, plus about $0.03 per reference. Krea’s report puts API pricing at $80 per 1,000 fast edits.

How many reference images can Krea 3 use?

Up to eight, referred to as @Image1 to @Image8 in the prompt.

Is Krea 3 better than Nano Banana Pro?

On the boards Krea cites, Krea 3 Ultra ranks #4 in image editing. Krea’s own benchmark shows it about three times faster and cheaper per edit. Which looks better depends on your images, so test both on your own work.

What is the difference between Krea 2 and Krea 3?

Krea 2 is an open text-to-image model you can run locally. Krea 3 builds on the Krea 2 pretrain but is a hosted model focused on multi-image editing.

Why does the same seed give different results?

Krea rewrites your prompt before generating, so the final prompt, and the image, changes between runs.

Sources

• Krea: Krea 3 technical report (3 November 2026)
• Krea: Krea 3 product and pricing page
• GitHub: ComfyUI PR #17067 (Krea 3 Image partner node)
• ComfyUI Wiki: Krea 3 technical report news